AI Agents Daily · Archive
AI Agents Daily archive, 7 issue
- 112026-11Agents Are Getting Wallets and Permissions—and Can't Patch Their Own HolesToday's agent beat is all about permissions and money. On one side, Microsoft disclosed a Lazarus backdoor that can wake dormant neurons in a model, letting agents with tool or terminal access execute malicious commands. On the other, Claude Code was found to have leaked 2,539 real credentials to GitHub through its settings files. On the product side, Rain is issuing agents payment accounts with spending caps, and Kata Send is offering cold-email infrastructure. Agents are being wired into real money and business workflows, but security boundaries are not keeping pace.
- 102026-11Thousand-Agent Orchestration Becomes the Selling Point as Security and Cost Emerge as Adoption BarriersAnthropic releases Fable/Mythos 5.2, showcasing thousand-subagent coordination and an operating system built in eight days, with API prices cut 20%; Meituan open-sources LongCat-2.5 with 1.6 trillion parameters, targeting long-horizon agent tasks; Baidu Dazi ranks first in three government and enterprise projects. Meanwhile, IEEE Spectrum discusses how zero-trust architecture can constrain agent behavior, and a V2EX user reports a suspected mass code-repository attack launched by AI agents, showing that verification and auditing are becoming the new bottleneck as agents scale.
- 092026-11Tencent Yuanqi Shuts Down as Agent Platforms Enter Consolidation PhaseTencent Yuanqi officially ceased service today, with all user data to be deleted after Jan 9, 2027; the company is directing users to migrate to Tencent Cloud ADP and WeChat Official Account smart replies. At the same time, Tencent showcased a more complete Agent lineup including WorkBuddy and Buddy One at its Digital Ecosystem Summit. Overseas, Apple's new CEO is pressing Siri to shift toward agents that autonomously execute tasks; Ant Group's Lingguang 2.0 managed 1.16 million tasks in a week, with users surpassing 20 million. The platform layer is contracting while the product layer is expanding.
- 082026-11Agents Move From Standalone Tools to Infrastructure: Collaboration, Interconnection and ComplianceGitHub unveils Rubicon, a new Git storage architecture that in internal testing supports 5,000 agents collaborating concurrently in a single repository and 7,200 pushes per second; Huawei, Alibaba and 18 other organizations jointly build agent interconnection infrastructure; Meta pushes Active Memory and Active Focus down to Muse and Meta AI. All three point the same way: agents are turning from point tools into infrastructure that must be hosted, interconnected and constrained.
- 072026-11Agents Move From Demo to Production: Capabilities, Incidents and EvaluationToday's main thread in the agent sector is deployment. Perplexity launched an agent built for sustained, complex work, pushing its product narrative from Q&A toward colleague-style collaboration. At the same time, an incident in which a Claude agent mistakenly deleted a law professor's entire year of lecture notes, and Chainalysis's disclosure that an attack involved AI agents, remind us that expanding capabilities bring operational risk. On the academic side, discussion centered on multi-agent information sharing, evaluation of personal service agents, and engineering platforms. Evaluation and governance are becoming preconditions for scaling agents.
- 062026-11AI Agents Shift From Capability Race to Governance and CollaborationToday's main thread is not a few more capable agents. It is that agents are starting to be treated as colleagues and infrastructure: Perplexity launched always-on Teammates to stake out work scenarios, Meta added real-time voice to Muse, and Microsoft Research used formal verification to constrain agent-written code. Meanwhile, gaps in agent identity and data provenance exposed on the enterprise side are becoming new procurement and M&A issues.
- 052026-11Agents Enter the Enterprise: Control and Traceability Become Prerequisites for DeploymentTwo threads point to the same problem: once agents move from demos into enterprise workflows, risk control becomes a prerequisite for deployment. MIT Technology Review proposes using data lineage to track data handoffs between multiple agents, avoiding chain errors that go unverified by humans; the other thread introduces "reverse game theory," which seeks to constrain strong model behavior through incentive design. Together they show that the competitive edge in agent productization is shifting from capability to controllability.